Robotics datasets: a primary-source reference map

DROID presents a dataset for robot manipulation. Start with its collection setup, available observations and actions, release instructions and evaluation material. Determine whether the documented hardware and tasks fit your question. The entry in this map does not imply independent verification of the project’s reported scale or performance.

Open X-Embodiment brings together robot-learning datasets across embodiments. Inspect constituent datasets, their documented formats and their individual terms. Aggregation does not make all component licences or task distributions identical. Keep the collection’s overall description separate from facts about one constituent release.

Ego4D focuses on first-person video and associated research tasks. Human observations can be relevant to embodied perception, but video is not automatically a robot-action trajectory. Check the access process, permitted uses, annotations and task definitions at the project. Avoid labelling human recordings as robot demonstrations unless a specific learning setup establishes that relationship.

Begin with the job you want the data to support. Is it robot action learning, first-person perception, evaluation or another task? Then inspect the action labels, embodiment, modalities and release rights. Do not rank projects by a mixed list of trajectories, clips, sensor-hours and unique hours. The values measure different things. A normalized comparison requires compatible definitions and a documented conversion basis, which this map does not claim to supply.

Download this reference map as CSV. The register includes the data category, primary URL, inspection question and reference date. It deliberately omits unsourced numeric rankings and blanket licence labels. Use the evidence checklist to extend it for a particular release. Report a stale link or correction through Feedback, with the original source and affected field. A future update should record its own date rather than pretending an older row is continuously verified.